It turns out to be a hallucination of memory retrieval; it's not really possible to see others' chat records. The characterization of a P0 level vulnerability is indeed an overstatement.

According to Beating reports, there are rumors that inputting specific tags into DeepSeek can reveal others' conversations and are characterized as P0 multi-tenant isolation failure. In fact, this is a phenomenon where the model generates seemingly real conversations based on memory and system prompts, rather than real-time retrieval of others' sessions. It is a common issue related to training data memory extraction and is not unique to DeepSeek. Academic research has shown that training data can be extracted through template tokens; the presence of the current date in conversations makes it difficult to prove they originate from others. Currently, to confirm actual user data leakage, concrete evidence is needed; there is no evidence pointing to the existence of other users.